Research On Visual Recognition Method Of Strawberry Picking Based On Deep Learning

With the rapid development of industry,agriculture and artificial intelligence,picking robots have been produced and applied as a kind of intelligent agricultural equipment.China is a large agricultural production country,the picking of various fruits and crops requires a lot of human and material resources,and the research of picking robots can effectively solve this problem.At present,the difficulties in the research of picking robots are mainly focused on target recognition and localization.A deep learning method is used to improve a strawberry fruit visual recognition network model to improve the accuracy and rate of the picking robot;a robot 3D hand-eye calibration method is improved to enhance the accuracy and speed of the hand-eye calibration,and then improve the efficiency of the picking robot.Finally,the designed method is experimentally validated.The main research content is as follows:(1)To study the basic theory of deep learning and derive the internal computational process of convolutional neural network to lay the foundation for designing the strawberry visual recognition detection model below.The current status of deep learning target detection algorithm research is analyzed,and the advantages and disadvantages of region generation-based target detection algorithm and regression-based target detection algorithm are compared,and the regression-based target detection method with faster detection rate is determined.(2)Introduce the camera imaging principle and the internal parameters of the camera through the camera imaging model,and derive the mathematical model for the conversion between the coordinate systems.The hardware facilities of the adopted depth camera are introduced,and the ranging principle is analyzed.The hand-eye calibration method is studied and the robot 3D hand-eye calibration method is improved to speed up the hand-eye calibration rate.(3)The YOLOv4-tiny network model was chosen as the benchmark,and the backbone network,neck network,and loss function of the model were improved to improve the accuracy and efficiency of the model for strawberry recognition.A dataset folder was created,the dataset was labeled,expanded,and the network model was trained.The tests showed that the strawberry recognition network model improved by the additive language has good detection accuracy and rate.(4)The robot strawberry vision grasping experimental platform was built,the camera was calibrated,and the improved hand-eye calibration method was experimentally verified.The results show that the improved hand-eye calibration method has good accuracy and rate.Strawberry grasping experiments were conducted using the improved strawberry recognition network model under three different strawberry placement methods,namely,flat,potted and hanging,and the experimental data of strawberry grasping were recorded.The experimental results show that the improved strawberry recognition network model has good accuracy and generalization.